Predicting sperm extraction in non-obstructive azoospermia patients

This is the authors' abstract. We don't add key points for this paper.

Sir, The management of male infertility due to non-obstructive azoospermia (NOA) with testicular sperm extraction (TESE) was a major advance in assisted reproduction (Schoysman et al., 1993). The largest case studies of well-defined NOA patients have reported sperm retrieval rates of around 50% (Verheyen et al., 2017). However, given the invasive nature of TESE, it is important to find predictive criteria for TESE outcomes in order to better counsel patients before surgery. To address this, Zeadna et al. (2020) present a ‘machine learning perspective’ that aims to predict the presence of sperm before surgery. This approach is fascinating, but a few points should be highlighted. The number of patients included in the study’s sample was relatively small (n = 119) and the authors excluded six patients (5%) from the statistical analyses because some data were missing. However, they did not provide any details about the missing data. Hence, it is not impossible that these exclusions unintentionally lead to patient selection bias. Furthermore, analysing the study population raises a crucial question. The authors define the population as ‘patients with NOA’. Presently, NOA is a clinical and paraclinical diagnosis of exclusion (absence of an obstructive cause) and is confirmed by testicular histology (impairment of spermatogenesis). Men with azoospermia, normal-size testes and normal serum FSH levels have normal spermatogenesis and are more likely to have obstructive azoospermia (OA), while men with a significant elevation in FSH have testicular failure, and thus testicular NOA. Furthermore, low ejaculate volume unrelated to hypogonadism or ejaculatory dysfunction is caused by obstruction (Practice Committee of the American Society for Reproductive Medicine in collaboration with the Society for Male Reproduction and Urology, 2018). Surprisingly, in this manuscript (Zeadna et al., 2020), 7.6% of patients with NOA showed normal spermatogenesis during the testicular histology examination. In addition, 25% of patients had a normal FSH level, and the mean testicular volume was 13 ml. Also, several patients had a below-normal volume of ejaculate. Finally, the sperm retrieval rates were much higher (65%) than those found in the literature for TESE (Verheyen et al., 2017). Given these different elements, it is not unlikely that some patients with OA or mixed azoospermia were included in the study, especially since sperm retrieval rates in patients with OA are much higher than in those with NOA. Thus, in our opinion, a predictive model, however powerful and computerised it may be, is only useful if the study population is strictly defined. In a logistic regression model, choosing the correct variables to enter into the model helps researchers to avoid two analytical pitfalls (Ranganathan et al., 2017). Firstly, including too many criteria could dilute true associations or identify spurious associations. Therefore, researchers must first perform univariate analyses and then include only those variables that meet a predefined significance level (P < 0.1–0.2 is commonly used). Thus, in the present study, only age, ethnicity and semen volume should have been included in the regression logistic model. Secondly, researchers must consider the scientific plausibility and clinical meaningfulness of the association. In this case, ethnicity was associated with the occurrence of sperm retrieval; however, this association is scientifically implausible and could be explained by its association with other variables, and therefore it should not have been entered into the logistic regression model. Additionally, the gradient boosted tree (GBT) model and its methodology was interesting; however, a good decision tree would separate the sampled population into a few groups that have low within-group variability but high between-group variability in outcome probability (Doupe et al., 2019). The GBT model presented included 36 groups and did not specify intra-group or inter-group variability. The chances of sperm extraction did not vary significantly between all groups (e.g. 0.97 vs 0.94), which suggests low inter-group variability for some branches of the decision tree. Finally, the GBT model displayed excellent sensitivity but low specificity (51%). Hence, one out of two patients for whom no sperm was found at the TESE had false hopes before surgery because the model came back positive. A few years ago, our team published a score that allowed a specificity higher than 80% (Boitrelle et al., 2011). In predictive models, it is important to consider patients’ counselling before surgery. An ideal predictive model would indeed avoid unnecessary TESE (i.e. failed surgeries), which requires high specificity to avoid false positives. In conclusion, more than 25 years after the first description of TESE, an ideal model for predicting the presence or absence of testicular spermatozoa before surgery remains elusive. None.